-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
683 lines (624 loc) · 33.8 KB
/
index.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
<!DOCTYPE html>
<html>
<head>
<title>Double Y: Building Extraction Generalization</title>
<link rel="icon" type="image/png" href="static/images/orbit.png">
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" -->
<meta name="description" content="Double Y: Building Extraction Generalization">
<meta property="og:title" content="Double Y: Building Extraction Generalization"/>
<meta property="og:description" content="Cross-City Building Instance Segmentation: From More Data to Diffusion-Augmentation"/>
<meta property="og:url" content="https://github.com/DoubleY-BEGC2024"/>
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="static/css/bulma-slider.min.css">
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="static/css/index.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script src="static/js/bulma-carousel.min.js"></script>
<script src="static/js/bulma-slider.min.js"></script>
<script src="static/js/index.js"></script>
</head>
<body>
<section class="hero">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">Cross-City Building Instance Segmentation: From More Data to Diffusion-Augmentation</h1>
<div class="is-size-5 publication-authors">
<!-- Paper authors -->
<span class="author-block">
<a href="https://www.linkedin.com/in/wongyijie/" target="_blank">Yi Jie WONG</a>,</span>
<span class="author-block">
<a href="https://www.linkedin.com/in/yinloonkhor/" target="_blank">Yin Loon KHOR</a></span>
</span>
</div>
<div class="is-size-5 publication-authors">
<span class="author-block"><strong>Group Name:</strong> Double-Y | <strong>Public
Leaderboard:</strong> 1st out of 68 Entrants <br>
<a href="https://www.kaggle.com/competitions/building-extraction-generalization-2024/overview">IEEE BigData Cup 2024: Building
Extraction Generalization Challenge</a>
</span>
<!-- <span class="eql-cntrb"><small><br><sup>*</sup>Indicates Equal Contribution</small></span> -->
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF -->
<span class="link-block">
<a href="static/pdfs/DoubleY Technical Report.pdf" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Technical Report</span>
</a>
</span>
<!-- PDF -->
<span class="link-block">
<a href="https://doi.org/10.36227/techrxiv.173091008.80781383/v1" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Full Paper</span>
</a>
</span>
<!-- Github link -->
<span class="link-block">
<a href="https://github.com/DoubleY-BEGC2024/OurSolution" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Source Code</span>
</a>
</span>
<!-- Best model -->
<span class="link-block">
<a href="https://www.dropbox.com/scl/fi/cdrl62i3mx9p82lqwpik5/yolov8m-seg_LasVegas.pt?rlkey=8ao7a5zz7xnqfd74deffprix2&st=m5kth3w0&dl=0"
target="_blank" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-dropbox"></i>
</span>
<span>Best Model</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Teaser GIF -->
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<img src="static/images/segmentations/output.gif" alt="Inference samples predicted by our trained model.">
<h2 class="subtitle">
Examples of Building Segmentation.
</h2>
</div>
</div>
</section>
<!-- End teaser GIF -->
<!-- Summary -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3">Summary</h2>
<div class="content has-text-justified">
<p> Deep learning has significantly advanced the field of building extraction from remote sensing images, providing robust solutions
for identifying and delineating building footprints. However, a major challenge persists in the form of domain adaptation, particularly
when addressing cross-city variations. The primary challenge lies in the significant differences in building appearances across cities,
influenced by variations in building shapes and environmental characteristics. Consequently, models trained on data from one city often
struggle to accurately identify buildings in another city. In this paper, we address this challenge from a data-centric perspective,
focusing on diversifying the training set. Our empirical results show that improving data diversity via open-source datasets and
diffusion augmentation significantly improved the performance of the segmentation model. Our baseline model, trained with no extra dataset,
only achieved a private F1 score of 0.663. On the other hand, our best model, trained with the additional Las Vegas building footprints
extracted from the Microsoft Building Footprint dataset, achieved a high private F1 score of 0.703. Surprisingly, we found that diffusion
augmentation helps improve our model score to 0.681 without requiring an extra dataset, which is higher than the baseline model. Finally,
we also experimented with the Non-Maximal Suppression (NMS) hyperparameter to improve the model’s performance in segmenting dense and
small objects, which gave us a high private F1 score of 0.897. Our source code and the pretrained models are publicly available at
https://github.com/DoubleY-BEGC2024/OurSolution.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- End paper abstract -->
<!-- Competition Overview -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3" style="white-space: nowrap;">Competition Overview</h4>
<div class="content has-text-justified">
<div style="text-align: justify;">
<p><strong>1. Objective:</strong> This competition embarks on this challenge by utilizing a building footprint dataset from the Tokyo area
as the primary training set, with plans to extend testing to other Japanese regions. This approach aims to inspire the development of models
with robust generalization capabilities, capable of overcoming the hurdles of automatic building footprint detection and extraction across
various landscapes. Overcoming this challenge signifies the creation of a novel approach for efficient, cost-effective, and precise building
footprint extraction at a national level with minimal regional data, showcasing its potential applicability worldwide.</p>
<p><strong>2. Mandatory Training Data:</strong> The training set data uses 0.3-meter Google earth satellite images complemented by meticulously manually annotated building outlines.
A total of 4717 images are provided, where all of them are extracted within Tokyo vicinity. The training data was divided into a training set
and a validation set with a ratio of 8:2.</p>
<p><strong>3. Mandatory Test Data:</strong> The imagery and building annotations for both test sets derive from the open-source Japanese 3D city model, the Plateau project (https://www.mlit.go.jp/plateau/)
enhanced with manual adjustments following visual inspection. All test images were randomly selected from 42 cities in Japan, but a balance of different types of
areas was maintained. A total of 250 images were taken from each region, totaling 1,000 images.</p>
</div>
</div>
<br><br>
</div>
</div>
<!-- Competition Overview -->
<!-- Methodology Overview -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3" style="white-space: nowrap;">Method 1: Additional Open-Sourced Dataset</h2>
<div class="content has-text-justified">
<div style="text-align: center;">
<img src="static/images/ROI cropped from Microsoft BF Dataset.png" alt="Microsoft BF Dataset" width="820">
<p class="caption" style="width: 100%; text-align: justify;">Figure 1: The region of interest for the building footprints
extracted from the Microsoft Building Footprint (BF) dataset. (a) Redmond, Washington. (b) Las Vegas, Nevada.
For simplicity's sake, we refer the former as Redmond dataset, and the latter as Las Vegas dataset.</p>
</div>
</div>
<br>
</div>
</div>
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3" style="white-space: nowrap;">Method 2: Diffusion Augmentation</h2>
<div class="content has-text-justified">
<div style="text-align: center;">
<img src="static/images/Segmentation Guided Diffusion.jpg" alt="Segmentation Guided Diffusion" width="820">
<p class="caption" style="width: 100%; text-align: justify;">Figure 2. The proposed diffusion augmentation pipeline.
(1) Use pretrained segmentation model to generate semantic segmentation.
(2) Refine the segmentation mask using the building polygon labels.
(3) Concatenate input image with the semantic mask.
(4) Train the segmentation-guided diffusion model using the concatenated inputs.</p>
</div>
</div>
<br>
</div>
</div>
<!-- End methodology overview-->
<!-- Model Selection -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3" style="white-space: nowrap;">Model Selection</h4>
<div class="content has-text-justified">
<div style="text-align: justify;">
<p>YOLOv8 series comes with several instance segmentation models, ranging from the smallest nano (n) variant to the
largest extra-large (x) variant. We performed several experiments to select the best YOLOv8 variant for our task,
considering both the F1 score and model complexity, as shown in Table I. Additionally, we compared the performance of
YOLOv8-based instance segmentation models with other state-of-the-art models, including YOLOv9, Mask R-CNN, and EfficientNet.
All models are trained for 50 epochs with 640 image size. During test time and submission, the confidence and NMS IoU
thresholds are set as 0.20 and 0.70, unless stated otherwise. We also tested the F1-score of the models with a confidence
threshold of 0.50, primarily to evaluate how confident the models are rather than for actual submission.
</p>
<table class="table is-bordered is-hoverable">
<thead>
<tr>
<th rowspan="2">Model</th>
<th rowspan="2">Pretrained Weights</th>
<th rowspan="2">Batch Size</th>
<th rowspan="2">Params (M)</th>
<th rowspan="2">FLOPs (G)</th>
<th colspan="2">Public F1-Score</th>
</tr>
<tr>
<th>Conf = 0.50</th>
<th>Conf = 0.20</th>
</tr>
</thead>
<tbody>
<tr>
<td>YOLOv8n-seg</td>
<td rowspan="4">DOTAv1 Aerial Detection</td>
<td>16</td>
<td>3.4</td>
<td>12.6</td>
<td>0.510</td>
<td>0.645</td>
</tr>
<tr>
<td>YOLOv8s-seg</td>
<td>16</td>
<td>11.8</td>
<td>42.6</td>
<td>0.535</td>
<td>0.654</td>
</tr>
<tr>
<td>YOLOv8m-seg</td>
<td>16</td>
<td>27.3</td>
<td>110.2</td>
<td>0.592</td>
<td>0.649</td>
</tr>
<tr>
<td>YOLOv8x-seg</td>
<td>8</td>
<td>71.8</td>
<td>344.1</td>
<td>0.579</td>
<td>0.627</td>
</tr>
<tr>
<td>YOLOv9c-seg</td>
<td rowspan="2">COCO Segmentation</td>
<td>4</td>
<td>27.9</td>
<td>159.4</td>
<td>0.476</td>
<td>0.577</td>
</tr>
<tr>
<td>Mask R-CNN (MPViT-Tiny)</td>
<td>4</td>
<td>17</td>
<td>196.0</td>
<td>-</td>
<td>0.596</td>
</tr>
<tr>
<td>EfficientNet-b0-YOLO-seg</td>
<td>ImageNet</td>
<td>4</td>
<td>6.4</td>
<td>12.5</td>
<td>-</td>
<td>0.560</td>
</tr>
</tbody>
</table>
<p>Our observations:</p>
<ol type="1" padding-left: 0;">
<li style="margin-bottom: 5px;">Generally, we observe that the F1 score increases when scaling up the model from
the smallest YOLOv8n-seg to the medium size YOLOv8m-seg. Notably, there is a significant jump in F1 score
from the YOLOv8s-seg to the YOLOv8m-seg when evaluated in confidence threshold of 0.50, with the score improving
from 0.535 to 0.592. </li>
<li style="margin-bottom: 5px;">Interestingly, the F1 score of YOLOv8m-seg is slightly lower than the smaller YOLOv8s-seg
when setting the confidence threshold to 0.20. This observation suggest that the m-variant still has a bigger room of
improvement compared to s-variant. </li>
<li style="margin-bottom: 5px;">Meanwhile, the largest YOLOv8x-seg variant has a lower F1 score than YOLOv8m-seg in
both confidence threshold 0.50 and 0.20. This suggests that further improvements in F1 score beyond the m-variant
may be minimal unless we enhance the quality of the training dataset or address generalizability issues.</li>
<li style="margin-bottom: 5px;"> We also tried training YOLOv9 instance segmentation model. Specficially, we chose YOLOv9c,
which corresponds to the m-variant of YOLOv8 (YOLOv8m-seg). However, we find that YOLOv9 is hard to train and slower
due to the high FLOPS, with no F1-score improvement.</li>
<li style="margin-bottom: 5px;"> Other than YOLO family, we also tried using Mask R-CNN with the MPViT backbone. However,
due to resource constraints, we were only able to test the smallest MPViT-Tiny for our Mask R-CNN backbone. The modified
Mask R-CNN has a slightly higher F1 score than YOLOv9c, but is inferior to all YOLOv8 variants we tried. It is not only
slow, but also not accurate.</li>
<li style="margin-bottom: 5px;"> Lastly, we tried replacing the YOLOv5 backbone with EfficientNet, which is a lightweight yet
effective CNN model designed for low-computational-power. We uses the smallest EfficientNet-b0 as the replacement backbone.
It is able to reach a considerably high F1 score (close to both YOLOv9c and Mask R-CNN) with significantly lower FLOPs.
However, the F1 score is still significantly lower than YOLOv8 series. We believe the performance could be further improved
by pretraining the EfficientNet with COCO segmentation, and scaling up the EfficientNet backbone.</li>
<li style="margin-bottom: 5px;"> In short, we prefered using YOLOv8 series due to its exceptional balance between speed and
accuracy. Specifically, we uses YOLOv8m-seg variant due to its high performance.</li>
</ol>
</div>
</div>
</div>
</div>
<!-- Model Selection -->
<!-- Dataset -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3" style="white-space: nowrap;">Dataset and Performance Improvement</h4>
<div class="content has-text-justified">
<div style="text-align: justify;">
<p>We experimented with the performance of YOLOv8m-seg by varying the training dataset, as shown in table below:
</p>
<table class="table is-bordered is-hoverable">
<tr>
<th>Setup</th>
<th>Dataset</th>
<th>Public F1 Score</th>
</tr>
<tr>
<td>A</td>
<td>BEGC 2024</td>
<td>0.649</td>
</tr>
<tr>
<td>B</td>
<td>BEGC 2024 + Redmond Dataset</td>
<td>0.660</td>
</tr>
<tr>
<td>C</td>
<td>BEGC 2024 + Las Vegas Dataset</td>
<td>0.686</td>
</tr>
<tr>
<td>D</td>
<td>BEGC 2024 + Diffusion Augmentation</td>
<td>0.672</td>
</tr>
<tr>
<td>E</td>
<td>BEGC 2024 + CutMix Dataset</td>
<td>0.650</td>
</tr>
</table>
<p>Our observations:</p>
<ul style="list-style-type: none; padding-left: 0;">
<li style="margin-bottom: 5px;"><strong>Setup A:</strong> 0.649 - This setup represents the performance baseline,
where YOLOv8m-seg was trained solely on the provided BEGC2024 training dataset. As expected, this setup
resulted in the lowest F1-score, likely due to the lack of diversity in the training data. </li>
<li style="margin-bottom: 5px;"><strong>Setup B:</strong> 0.660 - We trained YOLOv8m-seg using both the BEGC2024
training set and our Redmond dataset. This simple step of diversifying the training data led to a significant
increase in the F1-score, from 0.649 to 0.660. </li>
<li style="margin-bottom: 5px;"><strong>Setup C:</strong> 0.686 - Surprisingly, using the Las Vegas dataset resulted
in an even higher public F1-score of 0.686, as shown in Setup C. We believe the reason why the Las Vegas dataset
results in a greater improvement in F1 score is due to its greater semantic difference from the BEGC2024 training set,
which helps enhance the model's ability to generalize in the test set. </li>
<li style="margin-bottom: 5px;"><strong>Setup D:</strong> 0.672 - Surprisingly, the performance of the YOLOv8m-seg model
trained with the BEGC2024 dataset using diffusion augmentation resulted in a considerably high F1 score of 0.672.
This F1 score is even higher than that of Setup B, which was trained with the Redmond dataset. This observation
demonstrates that our diffusion augmentation method successfully created semantically different images that were
sufficient to diversify the BEGC2024 training set. </li>
<li style="margin-bottom: 5px;"><strong>Setup E:</strong> 0.650 - We also tried CutMix augmentation to diversify the
training dataset to improve generalization of our model. However, we found this method to be less effective,
achieving an F1 score of only 0.650, as shown in Setup E of Table III. The F1 score improvement was almost negligible
compared to our baseline in Setup A. We believe this is due to the lack of variation in building structures, as we
only changed the backgrounds. This highlights the importance of diversifying both building shapes and background
textures to improve the model's generalization. </li>
</ul>
</div>
</div>
</div>
</div>
<!-- Dataset -->
<!-- Kaggle Leaderboard -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3" style="white-space: nowrap;">Comparison with 2nd and 3rd Place Entrants</h4>
<div class="content has-text-justified">
<div style="text-align: justify;">
<p>We compare our solutions with the 2nd and 3rd place in the leaderboard:
</p>
<table class="table is-bordered is-hoverable">
<tr>
<th rowspan=2>Solution</th>
<th rowspan=2>FLOPS (G)</th>
<th colspan="2">F1-Score</th>
</tr>
<tr>
<td>Public</td>
<td>Private</td>
</tr>
<tr>
<td>YOLOv8m-seg + BEGC 2024</td>
<td rowspan=4>110.2</td>
<td>0.64926</td>
<td>0.66531</td>
</tr>
<tr>
<td>YOLOv8m-seg + BEGC 2024 + Redmond Dataset</td>
<td>0.65951</td>
<td>0.67133</td>
</tr>
<tr>
<td>YOLOv8m-seg + BEGC 2024 + Las Vegas Dataset</td>
<td>0.68627</td>
<td>0.70326</td>
</tr>
<tr>
<td>YOLOv8m-seg + BEGC 2024 + Diffusion Augmentation</td>
<td>0.67189</td>
<td>0.68096</td>
</tr>
<tr>
<td>2nd place (RTMDet-x + Alabama Buildings Segmentation Dataset)</td>
<td>141.7</td>
<td>0.6813</td>
<td>0.68453</td>
</tr>
<tr>
<td>3rd Place (Custom Mask-RCNN + No extra Dataset)</td>
<td>124.1</td>
<td>0.59314</td>
<td>0.60649</td>
</tr>
</table>
<p>Our observations:</p>
<ol type="1" padding-left: 0;">
<li style="margin-bottom: 5px;">Generally, using an additional dataset, whether it is an open-sourced dataset or a
synthetic dataset, helps improve the training of the model. </li>
<li style="margin-bottom: 5px;">However, you might sample high-quality or low-quality additional datasets from
open-sourced databases without careful engineering. For instance, using the Redmond dataset only slightly
improves the F1 score compared to using the BEGC 2024 dataset alone. On the other hand, using the Las Vegas dataset
significantly improves the F1 score, achieving the top F1 score among all methods.</li>
<li style="margin-bottom: 5px;">On the other hand, using our diffusion augmentation, we can generate a synthetic dataset
to train YOLOv8m-Seg without needing an additional dataset (which means no extra manual annotation is required).
Using BEGC2024 combined with the synthetic dataset, our YOLOv8m-Seg model achieved an F1 score that is significantly
higher than the baseline and close to our top-1 score (using the Las Vegas dataset) and the 2nd-place solution.</li>
<li style="margin-bottom: 5px;">Note that the 2nd-place solution uses a bigger model (higher FLOPs) with an additional
dataset to reach a high F1 score, whereas our diffusion augmentation pipeline allows our model (lower FLOPs) to
achieve a surprisingly close F1 score without an additional dataset.</li>
</ol>
</div>
</div>
</div>
</div>
<!-- Kaggle Leaderboard -->
<!-- NMS IoU Threshold -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3" style="white-space: nowrap;">Extra Trick: NMS IoU Threshold</h4>
<div class="content has-text-justified">
<div style="text-align: justify;">
<p>Non-maximal suppression (NMS) can be less effective at detecting small, densely packed objects, as it relies on
IoU to suppress overlapping bounding boxes. In scenarios involving small and dense objects, the bounding boxes
often overlap significantly, which can lead to the suppression of true positives. We can mitigate this issue by
increasing the IoU threshold in the NMS layer to prevent unnecessary reduction of bounding boxes. We experimented
by increasing the IoU threshold in the NMS layer of YOLOv8m-seg from the default 0.70 to 0.95, with increments of 0.05.
</p>
<table class="table is-bordered is-hoverable">
<thead>
<tr>
<th rowspan="2">Dataset</th>
<th colspan="6">Private F1 Score (using different NMS IoU Threshold)</th>
</tr>
<tr>
<th>0.70</th>
<th>0.75</th>
<th>0.80</th>
<th>0.85</th>
<th>0.90</th>
<th>0.95</th>
</tr>
</thead>
<tbody>
<tr>
<td>BEGC2024 + Redmond Dataset</td>
<td>0.672</td>
<td>0.677</td>
<td>-</td>
<td>-</td>
<td>0.748</td>
<td>0.866</td>
</tr>
<tr>
<td>BEGC2024 + Las Vegas Dataset</td>
<td>0.703</td>
<td>0.693</td>
<td>0.686</td>
<td>0.721</td>
<td>0.766</td>
<td>0.897</td>
</tr>
<tr>
<td>BEGC2024 + Diffusion Augmentation</td>
<td>0.681</td>
<td>-</td>
<td>0.694</td>
<td>0.711</td>
<td>0.751</td>
<td>0.887</td>
</tr>
</tbody>
</table>
<p>Our observations:</p>
<ol type="1" padding-left: 0;">
<li style="margin-bottom: 5px;">Generally, we found that IoU thresholds of 0.90 and 0.95 work best compared to
other threshold settings. </li>
<li style="margin-bottom: 5px;">Note that simply increasing the IoU threshold does not directly translate to
better performance, as it may lead to an increase in false positives that should have been suppressed by the
NMS layers. </li>
<li style="margin-bottom: 5px;">For instance, setting the IoU threshold between 0.75 and 0.80 is generally
worse than the default 0.70 threshold. </li>
<li style="margin-bottom: 5px;">Hence, our final submission is the YOLOv8m-seg model trained on the BEGC2024 and
Las Vegas datasets, with the IoU threshold for NMS set to 0.95. </li>
<li style="margin-bottom: 5px;">In future works, we consider trying more advanced NMS variation including Attention based NMS
and Density-based NMS to better mitigate this problem. </li>
</ol>
</div>
</div>
</div>
</div>
<!-- NMS IoU Threshold -->
<!-- Conclusion -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3" style="white-space: nowrap;">Key Takeaways</h4>
<div class="content has-text-justified">
<div style="text-align: justify;">
<p>
<strong>1. Dataset quality is what you need:</strong> There are 2 observations from our study. Firstly,
data diversity is important to mitigate the generalization challenge. For instance, Las Vegas dataset offers higher diversity
(i.e., desert backgrounds, different building shapes) as compared to the Redmond dataset, which is semantically more
similar to the provided BEGC2024 training set. Hence, the performance of our model trained with BEGC2024 + Las Vegas dataset
is better than BEGC2024 + Redmond dataset.
</p>
<p>
<strong>2. Diffusion Augmentation is label-efficient:</strong> Diffusion augmentaion is what you need if you do not have
extra dataset which is diverse enough from the original training set. For instance, the Redmond dataset is not as useful
as the Las Vegas dataset. However, it might be difficult and/or costly to find out the suitable extra dataset. On the other
hand, we do not need extra dataset to prepare our diffusion augmentation pipeline. Even better, BEGC2024 + Diffusion Augmentation
outperforms BEGC2024 + Redmond dataset, and also outperforms the 2nd and 3rd place entrants!
</p>
<p>
<strong>3. Start with a small model:</strong> We recommend starting with a smaller model. It is unwise to use a
larger model when dealing with a limited dataset, as it may lead to overfitting. Our empirical study agrees with
this hypothesis, as we failed to achieve a high mAP score using the biggest YOLOv8 version (YOLOv8x-seg). Given more time,
we would explore training YOLOv8x-seg with all the extra datasets we gathered, and also using our diffusion
augmentation pipeline.
</p>
</div>
</div>
<br><br>
</div>
</div>
<!-- Conclusion -->
<!-- Logo Acknowledgment -->
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h4 class="title is-3" style="white-space: nowrap;">Technological Stack</h4>
<div class="content has-text-justified">
<div style="text-align: justify;">
<p>
<a href="https://github.com/ultralytics/ultralytics" target="_blank"><img
src="static/icons/ultralyticsyolo-logo.svg" alt="ultralytics" style="width: 200px;"></a>
<a href="https://pytorch.org/" target="_blank"><img src="static/icons/pytorch-logo.svg" alt="pytorch"
style="width: 210px;"></a>
<a href="https://jupyter.org/" target="_blank"><img src="static/icons/jupyter-logo.png" alt="jupyter"
style="width: 200px;"></a>
<a href="https://www.python.org/" target="_blank"><img src="static/icons/python-logo.svg" alt="python"
style="width: 200px;"></a>
</p>
</div>
</div>
<br><br>
</div>
</div>
<!-- Logo Acknowledgment -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the <a href="https://github.com/eliahuhorwitz/Academic-project-page-template"
target="_blank">Academic Project Page Template</a> which was adopted from the <a
href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
You are free to borrow the of this website, we just ask that you link back to this page in the footer.
</p>
<p>
This website is licensed under a <a rel="license"
href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">Creative
Commons Attribution-ShareAlike 4.0 International License</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- Default Statcounter code for EY project website -->
<!--
<script type="text/javascript">
var sc_project = 12976265;
var sc_invisible = 1;
var sc_security = "c70be6f1";
</script>
<script type="text/javascript" src="https://www.statcounter.com/counter/counter.js" async></script>
<noscript>
<div class="statcounter"><a title="Web Analytics" href="https://statcounter.com/" target="_blank"><img
class="statcounter" src="https://c.statcounter.com/12976265/0/c70be6f1/1/" alt="Web Analytics"
referrerPolicy="no-referrer-when-downgrade"></a></div>
</noscript>
-->
<!-- End of Statcounter Code -->
</body>
</html>